自由空间光互连洗牌型神经网络模型

Free Space Optical Interconnected Perfect Shuffle Type of Neural Network Module

首先介绍一种洗牌型自由空间光互连多层全互连神经网络模型。该模型的高神经元／权重比可以极大地压缩神经网络的互连权矩阵IWM（interconnectweightmatrix）。对于具有N2个神经元的单层二维全互连神经网络的IWM为N2×N2，而洗牌型全互连神经网络的IWM仅为4N2log2N。另外，洗牌型全互连神经网络整齐、简单的结构方便了网络的综合，特别是网络隐单元的综合，并且十分适合于神经网络的光学实现。然后描述了采用数字光技术实现光互连的洗牌型神经网络的系统模型、关键芯片结构以及关键技术。本文提出的模型和方法使实现与人脑神经网络规模（104量级）相当的实用化自适应光电子全互连神经网络成为可能。

A perfect shuffle type of free space optical intcrconnected multi-layer globally connected nerual network module is firstly introduced. This model possesses high neuron to weight ratio, and can greatly shrink of the interconnect weight matrix(IWM).Besides, its simple and regular topology make it easy to synthesize the neural networks,especially for the hidden neuron, and suitable for the optical implementation. The digital optical interconnection technology based perfect shuffle type of neural network and the key chip and techniques are then described. Using the model and method introduced in this paper, it is possible to realize the large scale (104 neurons) optoelectronic fully connected neural networks.